Clustering and Classification of Satellite Imagery

Unsupervised Clustering

This section will use unsupervised clustering to split up given satellite images into several groups, with the goal of identifying meaningfull clusters that represent different geographic features such as water, farm land, forest or buildings.

## Loading required package: randomForest

## Warning in library(package, lib.loc = lib.loc, character.only = TRUE,
## logical.return = TRUE, : there is no package called 'randomForest'

• Load a raster file, which has 4 layers

rasterJ<-brick("C:\\Users\\Owner\\Documents\\Portugal\\Sat_Image_Classification\\data\\data\\img\\J_04AUG14112729-M2AS-000000137917_01_P001_etrs89.TIF")
#rasterJ<-brick("C:\\Users\\D059331\\Desktop\\DM GIC\\data\\img\\J_04AUG14112729-M2AS-000000137917_01_P001_etrs89.TIF")
rasterJ

## class       : RasterBrick 
## dimensions  : 4238, 4239, 17964882, 4  (nrow, ncol, ncell, nlayers)
## resolution  : 2.4, 2.4  (x, y)
## extent      : -105092.3, -94918.74, -50088.84, -39917.64  (xmin, xmax, ymin, ymax)
## coord. ref. : +proj=tmerc +lat_0=39.66825833333333 +lon_0=-8.133108333333334 +k=1 +x_0=0 +y_0=0 +ellps=GRS80 +units=m +no_defs 
## data source : C:\Users\Owner\Documents\Portugal\Sat_Image_Classification\data\data\img\J_04AUG14112729-M2AS-000000137917_01_P001_etrs89.TIF 
## names       : J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.1, J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.2, J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.3, J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.4 
## min values  :                                                  0,                                                  0,                                                  0,                                                  0 
## max values  :                                                255,                                                255,                                                255,                                                255

• Plot map

plotRGB(rasterJ, 3,2,1)

• choose a sector for clustering

ext <- extent(-104294.4, -102964.5, -43623.48, -42742.44 )
sector <- crop(rasterJ, ext)
plotRGB(sector, 3, 2, 1)

• add normdiff layer for better vegetation recognition

sector[[5]] <- normdiff(sector)

• plot each layer separately

plot(sector)

• check for NAs

summary(sector)

##         J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.1
## Min.                                                    23
## 1st Qu.                                                 31
## Median                                                  34
## 3rd Qu.                                                 40
## Max.                                                   176
## NA's                                                     0
##         J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.2
## Min.                                                    26
## 1st Qu.                                                 40
## Median                                                  50
## 3rd Qu.                                                 65
## Max.                                                   255
## NA's                                                     0
##         J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.3
## Min.                                                     8
## 1st Qu.                                                 17
## Median                                                  34
## 3rd Qu.                                                 51
## Max.                                                   194
## NA's                                                     0
##         J_04AUG14112729.M2AS.000000137917_01_P001_etrs89.4      values
## Min.                                                     4 -0.64885496
## 1st Qu.                                                 10 -0.20000000
## Median                                                  53  0.05645161
## 3rd Qu.                                                 70  0.15384615
## Max.                                                   191  0.76870748
## NA's                                                     0  0.00000000

• use only nir and ndvi layer for k-means

sector_mod <- brick(sector[[4]],sector[[5]])

• perform kmeans and plot result

performKMeans(sector_mod, 12)

K-Means doesnt seem to perform very well on this sector. Ocean and land arent clearly separated. Maybe this works better with fewer clusters (i.e. lower k).

performKMeans(sector_mod, 6)

Lets try instead a different sector, one without ocean.

ext <- extent(-101751, -100703.7, -47812.61, -46914.94)
landSector <- crop(rasterJ, ext)
landSector[[5]] <- normdiff(landSector)
plotRGB(landSector, 3, 2, 1)

landSector_mod <- brick(landSector[[4]],landSector[[5]])
performKMeans(landSector_mod, 12)

This is not too awesome. Try 6 clusters.

performKMeans(landSector_mod, 12)